Sd Mae Histopath Eval

sd-mae-histopath-eval

Self-distillation Augmented Masked Autoencoders for Histopathological Image Classification — Luo et al. (2022) (arXiv:2203.16983, 2022)

What this evaluates

Evaluates the ability of self-distillation augmented masked autoencoders to learn robust visual representations from histopathological images for downstream tasks like classification, segmentation, and detection, particularly in low-class or cross-domain settings.

Datasets

• PatchCamelyon (PCam) — total ?; splits: train (-1), val (-1), test (-1)
• NCT-CRC-HE (NCT) — total ?; splits: train (-1), val (-1), test (-1)
• MSIIvsMSS — total ?; splits: train (-1), val (-1), test (-1)
• MoNuSeg — total ?; splits: train (-1), val (-1), test (-1)
• Glas — total ?; splits: train (-1), val (-1), test (-1)
• NuCLS — total ?; splits: train (-1), val (-1), test (-1)

Metrics

• top-1 accuracy (primary) — range: percent
  • Fraction of correctly classified images out of the total number of images in the evaluation set.
• mIoU — range: percent
  • Mean Intersection over Union across all semantic classes, computed as the average of IoU per class.
• box mAP@0.5 — range: percent
  • Mean Average Precision for bounding box detection averaged over recall thresholds at an IoU threshold of 0.5.

Input / output format

Input: Histopathological image patches processed through a Vision Transformer (ViT) encoder with a 0.6 masking ratio during pre-training, and full images/patches during fine-tuning.

Output: Class probabilities for classification, pixel-wise segmentation masks for segmentation, and bounding boxes with class scores for object detection.

Scoring recipe

def compute_metrics(predictions, gold):
    acc = (predictions == gold).mean() * 100
    ious = intersection_over_union(predictions, gold)
    miou = ious.mean() * 100
    ap_scores = compute_ap(predictions, gold, iou_thresh=0.5)
    map50 = ap_scores.mean() * 100
    return {'top-1 accuracy': acc, 'mIoU': miou, 'box mAP@0.5': map50}

Common pitfalls

• Re-dividing datasets instead of strictly following the official train/val/test splits provided by the dataset creators.
• Using inconsistent fine-tuning protocols; the paper mandates a unified fine-tuning approach across all baselines for fair comparison.
• Ignoring high variance in contrastive learning baselines on low-class datasets, which requires multiple runs or careful standard deviation reporting.

Evidence (verbatim from paper)

Tab. 1 reports the top-1 accuracy of different methods on the two classification datasets. ... We report box mAP calculated at IoU threshold 0.5 on NuCLS [32].

Citation

@misc{luo2022selfdistillation,
  title={Self-distillation Augmented Masked Autoencoders for Histopathological Image Classification},
  author={Luo et al. (2022)},
  year={2022},
  note={arXiv:2203.16983}
}

• arXiv: 2203.16983